fix: scrub dangling lock-header leak + make benchmark stdout unicode-safe

Re-benchmark surfaced two issues:

1. Voice-reinforced adapters sometimes echo just the OPENING
   "=== PERMANENT BEHAVIORAL LOCKS (ABSOLUTE → NEVER VIOLATE) ===" header and
   trail off. The scrubber only matched complete open…close blocks, so the
   dangling header leaked into responses. Added pattern 2b: strip a lone opening
   LOCKS header to end-of-text (mirrors the existing CONSTRAINTS strip).

2. full_benchmark.py crashed mid-run printing a '→' to a cp1252 Windows console.
   Reconfigure stdout/stderr to utf-8 with errors='replace'.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: Jonathan Harrison

benchmarks/full_benchmark.py

@@ -14,6 +14,14 @@
 from pathlib import Path
 from datetime import datetime
 
+# Make stdout tolerant of non-cp1252 chars (responses may contain → — etc. on
+# Windows consoles). Without this, a single Unicode char crashes the whole run.
+try:
+    sys.stdout.reconfigure(encoding="utf-8", errors="replace")
+    sys.stderr.reconfigure(encoding="utf-8", errors="replace")
+except Exception:
+    pass
+
 # Ollama API
 import urllib.request
 

inference/codette_forge_bridge.py

@@ -682,6 +682,17 @@ def _scrub_leaked_directives(text: str) -> str:
             flags=_re.IGNORECASE | _re.DOTALL,
         )
 
+        # 2b. Strip a DANGLING opening LOCKS header (no closing marker) to end of
+        #     text. Voice-reinforced adapters sometimes echo just the opening
+        #     "=== PERMANENT BEHAVIORAL LOCKS (ABSOLUTE → NEVER VIOLATE) ===" and
+        #     trail off — pattern 2 only matches complete (open…close) blocks.
+        text = _re.sub(
+            r'={2,}\s*PERMANENT BEHAVIORAL LOCKS.*',
+            '',
+            text,
+            flags=_re.IGNORECASE | _re.DOTALL,
+        )
+
         # 3. Strip any remaining standalone LOCK lines (e.g. half-escaped blocks)
         text = _re.sub(
             r'\n?LOCK\s+\d+\s*[—\-]+\s+[A-Z][^\n]{0,300}',